AI and Workflow Automation: Time and Money Back to Your Team

An AI agent is a program that uses a language model, such as Claude or ChatGPT, to do a series of tasks, not only answer a question: read an email, look up information in your tools, prepare a reply. In a company, it prepares; a person checks and decides.

Start with one precise, recurring task, not with the tool. Write down the steps, the special cases and what counts as a good result. Then give the agent access to only the tools it needs, with Claude, ChatGPT or a platform such as Make. Test it on real cases, and keep a person who reviews.

Work that always follows the same steps: re-keying orders or data between two tools, chasing quotes and invoices, sorting emails, standard quotes, booking appointments, monthly reports, filing documents. Work that needs judgement stays with a person.

Count first: which tasks come back every week, how many hours they take, and whether they always follow the same steps. Start with the one that gives back the most hours, in the tools you already have. One flow at a time, with an owner and an off switch. The task map at the top of this page does this count in 3 minutes.

It means giving a flow a series of steps that always repeats the same way: an order that goes from the form to the CRM and then to accounting, a reminder that goes out on the planned day. The flow follows written rules. It saves time on repetitive work, and it leaves the decisions to a person.

It is a flow that contains one step that fixed rules cannot do: read a free-text email, extract the data from an invoice, write a first draft. AI does that step; the rest of the flow stays ordinary. A person reads everything that goes to a client.

With one task, frequent and well known, not with a big project. Measure the time it takes today, try AI on ten real cases, and compare. Then decide where your data is and which provider processes it. If the gain is real, move on to the next task.

A flow can read invoices that arrive by email: amount, date, supplier, VAT. The data then goes into your accounting, and a person approves before payment. Many accounting programs already offer this reading; if yours does not, a flow does it between your mailbox and your accounting.

Each quote you send gets a follow-up date. With no reply by that date, a reminder goes out, or a draft waits for your approval. Same logic for overdue invoices. The tone stays yours: you write the texts once.

Sort first: quote requests, questions, invoices, advertising. A flow can classify incoming emails and prepare a reply from your texts and your data. The reply goes out when a person has read it, until you decide otherwise.

Zapier is an online service that connects your apps without code: when something happens in one tool, an action happens in another. It has a free plan limited in tasks per month; beyond that, plans are paid and priced on volume.

Both connect your tools without code. Zapier is the easiest to learn and has the most connectors. Make is more visual for flows with several steps and is often cheaper at the same volume. For a critical or sensitive flow, a small custom program is sometimes safer and cheaper over time.

It depends on the task. To write, summarise and reason about texts, Claude, ChatGPT and Gemini are the most used. To keep data in Switzerland, Apertus, the open model from EPFL, ETH Zurich and CSCS released in September 2025, is an option. The right choice also depends on where your data is processed.

Connect the report to its sources (accounting, CRM, spreadsheet) and set up the template once. The report then fills itself in every month and goes out as a PDF to the right people. On the first day of the month, it is ready.

Only with your consent, and only what is necessary. Before we start, I tell you which provider is used and where it processes the data, preferably in Switzerland or the EU. The provider must not train its models on your data.

Each flow stops and alerts someone when it fails. A person reads everything a client sees. Each flow has a log, an owner on your side and an off switch.

It is often the right answer, and I say so when it is. Custom software is worth it when your way of working is your advantage, or when the off-the-shelf tool costs more in adjustments than a small tool made for you.

I do, if you want, under a maintenance contract. Otherwise, the documentation is enough for another developer.

On your own hosting where possible, in Switzerland or Europe. I tell you where before we start.

It depends on the tool. A small, useful first version first; the rest is added in steps, each one costed. The duration is written in the offer.

On a quote, after the thirty-minute call. The task map gives the time the tool would give back: that is the upper limit of what it should cost.